xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x16c2-minmax-rndnu-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 4);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (4 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53   if XNN_UNPREDICTABLE(mr != 4) {
54     c3 = c2;
55   }
56 
57   do {
58     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
61     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62     int32x4_t vacc1x0123 = vacc0x0123;
63     int32x4_t vacc1x4567 = vacc0x4567;
64     int32x4_t vacc1x89AB = vacc0x89AB;
65     int32x4_t vacc1xCDEF = vacc0xCDEF;
66     int32x4_t vacc2x0123 = vacc0x0123;
67     int32x4_t vacc2x4567 = vacc0x4567;
68     int32x4_t vacc2x89AB = vacc0x89AB;
69     int32x4_t vacc2xCDEF = vacc0xCDEF;
70     int32x4_t vacc3x0123 = vacc0x0123;
71     int32x4_t vacc3x4567 = vacc0x4567;
72     int32x4_t vacc3x89AB = vacc0x89AB;
73     int32x4_t vacc3xCDEF = vacc0xCDEF;
74 
75     size_t p = ks;
76     do {
77       const int8_t* restrict a0 = a[0];
78       if XNN_UNPREDICTABLE(a0 != zero) {
79         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80       }
81       const int8_t* restrict a1 = a[1];
82       if XNN_UNPREDICTABLE(a1 != zero) {
83         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84       }
85       const int8_t* restrict a2 = a[2];
86       if XNN_UNPREDICTABLE(a2 != zero) {
87         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88       }
89       const int8_t* restrict a3 = a[3];
90       if XNN_UNPREDICTABLE(a3 != zero) {
91         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
92       }
93       a += 4;
94 
95       size_t k = kc;
96 
97       while (k >= 16 * sizeof(int8_t)) {
98         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
99         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
100         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
101         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
102         const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
103         const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
104         const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
105         const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
106         const int16x4x2_t va20x0 = vld2_dup_s16((const void*)a2);
107         const int16x4x2_t va21x0 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
108         const int16x4x2_t va20x1 = vld2_dup_s16((const void*)a2);
109         const int16x4x2_t va21x1 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
110         const int16x4x2_t va30x0 = vld2_dup_s16((const void*)a3);
111         const int16x4x2_t va31x0 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
112         const int16x4x2_t va30x1 = vld2_dup_s16((const void*)a3);
113         const int16x4x2_t va31x1 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
114 
115         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 
132         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
133         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
134         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
135         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
136         const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0.val[0]);
137         const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1.val[0]);
138         const int8x8_t va3c0x0 = vreinterpret_s8_s16(va30x0.val[0]);
139         const int8x8_t va3c0x1 = vreinterpret_s8_s16(va30x1.val[0]);
140 
141         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
142         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
143         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
144         int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
145         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
147         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
148         vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
149         vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
150         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
151         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
152         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
153         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
154         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
155         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
156         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
157         int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
158         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
160         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
161         vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
162         vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
163         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
164         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
165         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
166         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
167         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
168         int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
169         int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0);
170         int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3c0x0);
171         const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172         vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
173         vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
174         vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2c0x1);
175         vprod3x89ABc0 = vmlal_s8(vprod3x89ABc0, vb89ABc0x1, va3c0x1);
176         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
177         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
178         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
179         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
180         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
181         int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
182         int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0);
183         int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3c0x0);
184         const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185         vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
186         vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
187         vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2c0x1);
188         vprod3xCDEFc0 = vmlal_s8(vprod3xCDEFc0, vbCDEFc0x1, va3c0x1);
189         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
190         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
191         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
192         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
193         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
194         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
195         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
196         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
197         const int8x8_t va2c1x0 = vreinterpret_s8_s16(va20x0.val[1]);
198         const int8x8_t va2c1x1 = vreinterpret_s8_s16(va20x1.val[1]);
199         const int8x8_t va3c1x0 = vreinterpret_s8_s16(va30x0.val[1]);
200         const int8x8_t va3c1x1 = vreinterpret_s8_s16(va30x1.val[1]);
201 
202         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
203         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
204         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
205         int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
206         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
208         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
209         vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
210         vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
211         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
212         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
213         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
214         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
215         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
216         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
217         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
218         int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
219         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
221         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
222         vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
223         vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
224         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
225         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
226         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
227         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
228         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
229         int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
230         int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2c1x0);
231         int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0);
232         const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233         vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
234         vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
235         vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2c1x1);
236         vprod3x89ABc1 = vmlal_s8(vprod3x89ABc1, vb89ABc1x1, va3c1x1);
237         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
238         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
239         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
240         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
241         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
242         int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
243         int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2c1x0);
244         int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0);
245         const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
247         vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
248         vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2c1x1);
249         vprod3xCDEFc1 = vmlal_s8(vprod3xCDEFc1, vbCDEFc1x1, va3c1x1);
250         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
251         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
252         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
253         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
254         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
255         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
256         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
257         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
258         const int8x8_t va2c2x0 = vreinterpret_s8_s16(va21x0.val[0]);
259         const int8x8_t va2c2x1 = vreinterpret_s8_s16(va21x1.val[0]);
260         const int8x8_t va3c2x0 = vreinterpret_s8_s16(va31x0.val[0]);
261         const int8x8_t va3c2x1 = vreinterpret_s8_s16(va31x1.val[0]);
262 
263         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
264         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
265         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
266         int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
267         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
269         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
270         vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
271         vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
272         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
273         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
274         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
275         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
276         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
277         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
278         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
279         int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
280         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
282         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
283         vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
284         vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
285         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
286         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
287         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
288         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
289         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
290         int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
291         int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2c2x0);
292         int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3c2x0);
293         const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294         vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
295         vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
296         vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2c2x1);
297         vprod3x89ABc2 = vmlal_s8(vprod3x89ABc2, vb89ABc2x1, va3c2x1);
298         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
299         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
300         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
301         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
302         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
303         int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
304         int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2c2x0);
305         int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3c2x0);
306         const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
307         vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
308         vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
309         vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2c2x1);
310         vprod3xCDEFc2 = vmlal_s8(vprod3xCDEFc2, vbCDEFc2x1, va3c2x1);
311         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
312         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
313         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
314         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
315         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
316         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
317         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
318         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
319         const int8x8_t va2c3x0 = vreinterpret_s8_s16(va21x0.val[1]);
320         const int8x8_t va2c3x1 = vreinterpret_s8_s16(va21x1.val[1]);
321         const int8x8_t va3c3x0 = vreinterpret_s8_s16(va31x0.val[1]);
322         const int8x8_t va3c3x1 = vreinterpret_s8_s16(va31x1.val[1]);
323 
324         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
325         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
326         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
327         int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
328         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
330         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
331         vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
332         vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
333         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
334         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
335         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
336         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
337         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
338         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
339         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
340         int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
341         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
342         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
343         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
344         vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
345         vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
346         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
347         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
348         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
349         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
350         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
351         int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
352         int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0);
353         int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3c3x0);
354         const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
355         vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
356         vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
357         vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2c3x1);
358         vprod3x89ABc3 = vmlal_s8(vprod3x89ABc3, vb89ABc3x1, va3c3x1);
359         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
360         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
361         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
362         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
363         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
364         int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
365         int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2c3x0);
366         int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3c3x0);
367         const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
368         vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
369         vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
370         vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2c3x1);
371         vprod3xCDEFc3 = vmlal_s8(vprod3xCDEFc3, vbCDEFc3x1, va3c3x1);
372         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
373         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
374         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
375         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
376 
377         k -= 16 * sizeof(int8_t);
378       }
379 
380       if (k >= 8 * sizeof(int8_t)) {
381         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
382         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
383         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
384         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
385         const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
386         const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
387         const int16x4x2_t va30 = vld2_dup_s16((const void*)a3);
388         const int16x4x2_t va31 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
389 
390         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
391         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
392         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
393         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
394         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
395         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
396         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
397         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
398         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
399         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
400         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
401         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
402         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
403         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
404         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
405         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
406 
407         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
408         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
409         const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
410         const int8x8_t va3c0 = vreinterpret_s8_s16(va30.val[0]);
411 
412         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
413         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
414         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
415         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
416         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
417         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
418         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
419         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
420         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
421         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
422         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
423         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
424         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
425         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
426         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
427         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
428         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
429         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
430         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
431         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
432         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
433         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
434         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
435         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
436         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
437         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
438         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
439         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
440         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
441         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
442         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
443         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
444         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
445         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
446         const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
447         const int8x8_t va3c1 = vreinterpret_s8_s16(va30.val[1]);
448 
449         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
450         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
451         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
452         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
453         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
454         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
455         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
456         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
457         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
458         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
459         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
460         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
461         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
462         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
463         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
464         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
465         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
466         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
467         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
468         const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
469         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
470         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
471         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
472         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
473         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
474         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
475         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
476         const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
477         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
478         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
479         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
480         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
481         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
482         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
483         const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
484         const int8x8_t va3c2 = vreinterpret_s8_s16(va31.val[0]);
485 
486         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
487         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
488         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
489         const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
490         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
491         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
492         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
493         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
494         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
495         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
496         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
497         const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
498         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
499         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
500         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
501         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
502         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
503         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
504         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
505         const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
506         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
507         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
508         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
509         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
510         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
511         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
512         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
513         const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
514         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
515         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
516         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
517         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
518         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
519         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
520         const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
521         const int8x8_t va3c3 = vreinterpret_s8_s16(va31.val[1]);
522 
523         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
524         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
525         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
526         const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
527         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
528         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
529         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
530         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
531         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
532         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
533         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
534         const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
535         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
536         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
537         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
538         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
539         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
540         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
541         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
542         const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
543         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
544         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
545         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
546         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
547         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
548         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
549         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
550         const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
551         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
552         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
553         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
554         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
555 
556         k -= 8 * sizeof(int8_t);
557       }
558 
559       if XNN_UNLIKELY(k != 0) {
560         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
561         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
562         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
563         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
564 
565         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
566         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
567         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
568         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
569 
570         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
571         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
572         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
573         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
574         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
575         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
576         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
577         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
578         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
579         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
580         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
581         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
582         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
583         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
584         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
585         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
586         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
587         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
588         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
589         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
590         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
591         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
592         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
593         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
594         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
595         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
596         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
597         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
598         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
599         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
600         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
601         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
602         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
603         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
604         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
605         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
606 
607         if (k > 2 * sizeof(int8_t)) {
608           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
609           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
610           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
611           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
612 
613           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
614           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
615           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
616           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
617           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
618           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
619           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
620           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
621           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
622           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
623           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
624           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
625           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
626           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
627           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
628           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
629           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
630           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
631           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
632           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
633           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
634           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
635           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
636           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
637           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
638           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
639           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
640           const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
641           const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
642           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
643           const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
644           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
645           const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
646           vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
647           const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
648           vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
649 
650           if (k > 4 * sizeof(int8_t)) {
651             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
652             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
653             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
654             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
655 
656             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
657             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
658             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
659             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
660             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
661             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
662             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
663             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
664             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
665             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
666             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
667             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
668             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
669             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
670             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
671             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
672             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
673             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
674             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
675             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
676             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
677             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
678             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
679             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
680             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
681             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
682             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
683             const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
684             const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
685             vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
686             const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
687             vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
688             const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
689             vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
690             const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
691             vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
692           }
693         }
694       }
695       p -= 4 * sizeof(void*);
696     } while (p != 0);
697 
698     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
699     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
700     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
701 
702     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
703     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
704     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
705     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
706     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
707     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
708     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
709     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
710     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
711     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
712     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
713     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
714     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
715     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
716     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
717     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
718 
719     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
720     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
721     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
722     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
723     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
724     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
725     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
726     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
727     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
728     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
729     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
730     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
731     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
732     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
733     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
734     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
735 
736     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
737     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
738     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
739     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
740     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
741     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
742     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
743     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
744     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
745     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
746     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
747     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
748     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
749     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
750     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
751     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
752 
753     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
754 #if XNN_ARCH_ARM64
755     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
756     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
757     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
758     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
759     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
760     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
761     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
762     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
763 
764     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
765     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
766     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
767     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
768     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
769     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
770     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
771     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
772 
773     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
774     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
775     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
776     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
777 #else
778     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
779     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
780     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
781     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
782     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
783     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
784     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
785     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
786 
787     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
788     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
789     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
790     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
791     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
792     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
793     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
794     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
795 
796     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
797     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
798     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
799     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
800 #endif
801 
802     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
803     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
804     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
805     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
806     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
807 
808     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
809     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
810     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
811     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
812     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
813 
814     if (nc >= 16) {
815       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
816       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
817       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
818       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
819 
820       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
821       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
822       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
823       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
824 
825       a = (const int8_t**restrict) ((uintptr_t) a - ks);
826 
827       nc -= 16;
828     } else {
829       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
830       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
831       if (nc & 8) {
832         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
833         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
834         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
835         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
836         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
837         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
838       }
839       if (nc & 4) {
840         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
841         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
842         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
843         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
844         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
845         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
846       }
847       if (nc & 2) {
848         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
849         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
850         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
851         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
852         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
853         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
854       }
855       if (nc & 1) {
856         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
857         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
858         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
859         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
860       }
861 
862       nc = 0;
863     }
864   } while (nc != 0);
865 }
866